Qdrant MCP Server 🔌 MCP server Open source

What it does

mcp-server-qdrant is Qdrant's official MCP server. It gives an agent a semantic memory layer on top of the Qdrant vector database. It is deliberately small: the agent stores a piece of text with optional metadata, the server embeds it with a local FastEmbed model and writes it to a collection, and later the agent retrieves the closest matches with a natural-language query. You can change the tool descriptions through environment variables to repurpose it. The README shows how to turn it into a snippet library, where the agent saves reusable code with a plain-English description and searches it later, and gives setups for Claude Code, Cursor and Windsurf. The repository calls itself an example implementation, so expect a simple building block rather than a full retrieval system.

Tools it gives your agent

• qdrant-store: embed and save information, with optional JSON metadata, to a collection
• qdrant-find: semantic search over a collection, returning up to QDRANT_SEARCH_LIMIT results, 10 by default

Works with

It runs locally with uvx or Docker. It uses stdio by default, and SSE or Streamable HTTP are available for remote or shared use. The README covers Claude Desktop, Claude Code, Cursor, Windsurf and VS Code, and any MCP client will work. It connects to a Qdrant server or Qdrant Cloud through QDRANT_URL, or to an embedded on-disk database through QDRANT_LOCAL_PATH.

How to connect

The README's basic run command (Claude Code users can wrap the same variables in claude mcp add ... -- uvx mcp-server-qdrant):

QDRANT_URL="http://localhost:6333" \
COLLECTION_NAME="my-collection" \
EMBEDDING_MODEL="sentence-transformers/all-MiniLM-L6-v2" \
uvx mcp-server-qdrant

Cost and limits

The server is free under Apache-2.0. Self-hosted Qdrant and local-path mode cost nothing, while Qdrant Cloud has its own pricing. Only FastEmbed models are supported, and embeddings run on your machine. There are no tools to delete, update or list entries, so you manage collections through Qdrant itself.

Safety notes

Set QDRANT_READ_ONLY=true to remove the store tool and give the agent search only. Keep QDRANT_API_KEY in your client's secret inputs, not in a committed config file. The HTTP transports bind to 127.0.0.1 by default. The Docker example binds to 0.0.0.0, so restrict network access if you use it. Stored content comes back into the agent's context, so do not store untrusted text.